Slides - Grid`5000

Transcription

Slides - Grid`5000
Building an energy proportional datacenter
with an heterogeneous set of machines
Violaine V ILLEBONNET
Laurent L EFEVRE – Inria Avalon E NS Lyon
Jean-Marc P IERSON , Georges D A C OSTA, Patricia S TOLF – I RIT Toulouse
Grid'5000 School
February 2 nd 2016
Looking for Energy Proportionality
« The case for Energy-Proportional Computing »
L. A. Barroso and U. Hölzle, IEEE Computer , 2007
OBSERVATIONS:
●
High idle consumption
Can be up to 50 % of peak power
●
Average server load
between 10 and 50 %
Most inefficient region
●
Uptime Institute, 2015
●
30% of US servers are « comatose »
« those that have not delivered information
or computing services in six months or more »
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
3
« Big Medium Little » datacenter infrastructure
HETEROGENEOUS INFRASTRUCTURE composed of machines
with different characteristics in terms of performance and energy consumption
BIG
•
Classical servers
→ Only used at their most energy
efficient region
•
Low power processors
→ Reduce static costs
MEDIUM
LITTLE
TECHNICAL CHALLENGES :
•
Migrations between heterogeneous architectures (ARM and x86)
•
Placement: Dynamically find the most suitable combinations of machines
•
Reconfiguration: Power On/Off machines at the right time
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
5
Overview of BML Framework
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
6
Chosen hardware to profile
FROM GRID'5000:
●
Paravance (Rennes)
x86 Intel Xeon E5-2630v3 (2 x 8 cores)
●
Taurus (Lyon)
x86 Intel Xeon E5-2630 (2 x 6 cores)
●
Graphene (Nancy)
x86 Intel Xeon X3440 (1 x 4 cores)
AND FOR LOW POWER:
●
Samsung Chromebook
ARM Cortex-A15 (1 x 2 cores)
●
Raspberry Pi 2
ARM Cortex-A7 (1 x 4 cores)
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
8
Experiments for machine profiling
APPLICATION PERFORMANCE and POWER CONSUMPTION:
●
Use-case application: stateless web server
●
(Performance metric: requests/s)
→ Lighttpd
Benchmark to generate load → Siege
and gather performance metrics
●
Power monitoring
Fetch 1 Watt value per sec, average on benchmark period
ON/OFF DURATION and ENERGY CONSUMPTION:
●
Switch off, until (nearly) no power consumption
●
Switch on, until machine answers to ping
●
Power monitoring:
Fetch 1 Watt value per sec, sum to Joules
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
9
Experiments for machine profiling
APPLICATION PERFORMANCE and POWER CONSUMPTION:
●
Use-case application: stateless web server
(Performance metric: requests/s)
●
Benchmark to generate load → Lighttpd
G RID'5000 TOOLS:
2 machines with my
custom environment
➢
kadeploy3
➢
kwapi ➢
kapower3
➢
kwapi
→ Siege
and gather performance metrics
●
Power monitoring
Fetch 1 Watt value per sec, average on benchmark period
ON/OFF DURATION and ENERGY CONSUMPTION:
●
Switch off, until (nearly) no power consumption
●
Switch on, until machine answers to ping
●
Power monitoring:
Fetch 1 Watt value per sec, sum to Joules
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
10
Simulations
MOTIVATIONS:
No heterogeneous datacenter at disposal
Evaluate the achievable gains while modifying input parameters easily
TOOL:
Simulator developed in Python
Inputs: Machine profiles in Json and Trace file
General functioning (for presented results):
- Perfect future load prediction (max or average on a sliding window)
- Compute BML ideal combination for predicted load
- Perform the reconfiguration towards the new combination
- Once reconfiguration done, loop
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
15
CONCLUSION ON GRID'5000 FOR MY EXPERIMENTS
●
Easy to profile performance and energy of several architectures
●
It would be great to have more heterogeneity in the platform :)
ON-GOING/FUTURE WORK
•
Enhance reconfiguration decisions
•
Evaluate impact of load prediction errors
•
Validation with other use-case applications
•
Apply to limited datacenter infrastructure
violaine.villebonnet@inria.fr
Grid'5000 Winter School - February 2nd 2016
Violaine Villebonnet
18